Kimi K3.1 Pro: An Open Model Within a Point of Claude Opus 5.5?
Moonshot's Kimi K3.1 Pro lands within a point of Claude Opus 5.5 on Terminal-Bench, costs a third less than K3, and promises open weights by Nov 4.
Moonshot AI just released Kimi K3.1 Pro, and on its own numbers the model lands within a single point of Claude Opus 5.5 on Terminal-Bench 4.0 while costing a third less than its predecessor. The kicker: Moonshot says the full weights will be public by November 4.
Here's what you need to know:
Kimi K3.1 Pro launched on October 23, 2026 on Kimi.ai, Kimi Work, Kimi Code and the Kimi API, alongside a Highspeed variant that Moonshot says is 2-3x faster.
Moonshot reports 65.7% on Terminal-Bench 4.0, versus 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra.
API pricing is $2 per million input tokens and $10 per million output, down from $3 and $15 for Kimi K3.
It keeps a 1-million-token context window and, per Yotta Labs, K3's architecture of 2.8 trillion total and 104 billion active parameters.
The open weights are due by November 4, but the license hasn't been announced.
Moonshot's own benchmark table: close to Opus 5.5 on terminal tasks, ahead on long-horizon coding.
What Happened
In its official launch post, Moonshot calls Kimi K3.1 Pro "the first model in our new K3.1 series and our most capable 3T-class open-weight model to date." It's a post-trained version of Kimi K3 with "substantially larger post-training compute," using reinforcement learning across hundreds of domains.
The headline numbers are about coding agents. Moonshot says K3.1 Pro scores 65.7% on Terminal-Bench 4.0 and 61.2% on its internal Kimi Code Bench 3.0 Long-horizon, up from 55.95% for K3. On BenchCAD, which turns multi-view renderings into CAD code, it reports 92.7% with tools, behind Claude Opus 5.5 (96.2%) and GPT-6 Astra (95.9%).
Pricing is where it gets interesting. According to Kimi's pricing page, kimi-k3.1-pro costs $2.00 per million input tokens, $10.00 per million output and $0.20 per million cached input, with flat pricing across the whole 1,048,576-token window. The Highspeed model doubles those rates. Yotta Labs' breakdown notes that Kimi K3 was $3.00 / $0.30 / $15.00, so every rate fell by a third.
Moonshot also claims a big drop in made-up facts: on its internal Kimi Hallu Bench Hard stress test, the claim-level hallucination rate fell from 23.4% for K3 to 8.3%. And it says it removed cyber-related data from training before the open-weight release. On its in-house exploit benchmark, a checkpoint without that filtering solved 90.5% of tasks at pass@3, while the final model solved 81.0%.
Why It Matters
Open-weight models used to trail the best closed models by a generation. If Moonshot's numbers hold up, K3.1 Pro narrows that to a rounding error on at least one serious coding-agent benchmark, at a fraction of closed-model prices. Yotta Labs estimates a workload of 10 million input and 2 million output tokens a day, with no caching, costs about $1,200 a month on K3.1 Pro, against $6,000 on GPT-6 Astra.
There's an important caveat. Moonshot ran K3.1 Pro with the Claude Code harness and took other models' scores from their own reports or public leaderboards, so the table mixes sources. Early independent results are encouraging, though. In Plotly's data analytics benchmark, K3.1 Pro passed 42 of 43 questions, up from 35 for Kimi K3, tying for third place on a leaderboard of 34 models.
It also keeps up the pressure from Chinese labs on price, following moves like DeepSeek's V4.1-Flash update and its $0.75 search API.
What It Means for Developers
You can try it today without changing much code. Yotta Labs says the Kimi API is OpenAI-compatible, so switching an existing client is mostly a base URL and a model string.
How a request flows to K3.1 Pro, and where prompt caching saves money.
Benchmark it on your own tasks. Vendor scores aren't your codebase. Run your agent evals against
kimi-k3.1-probefore switching anything in production.Design for caching. Cached input is a tenth of the regular price. Plotly found 92% of the tokens it sent were cache hits, so stable system prompts and repo context pay off.
Watch output costs and latency. Plotly measured a median of 2.5 minutes per question and about 16,500 output tokens, mostly reasoning. Output made up $7.10 of its $13.47 bill.
Use Highspeed only where latency matters. It's the same model at double the price, served 2-3x faster according to Moonshot.
Plan self-hosting realistically. Plotly notes K3's weights were about 1.6 TB, so this is GPU-cluster territory, not a laptop model.
What's Next
The big date is November 4, when Moonshot says the full weights will drop. The license is the open question: Kimi K3 shipped under a custom license that, according to Plotly, asks the largest commercial inference providers to share revenue. Moonshot hasn't said whether K3.1 Pro will follow it.
Expect independent leaderboards to fill in over the next week. OpenRouter already lists the model at the same $2 and $10 rates, which makes side-by-side testing easy. For comparison shopping, see our guide to Gemini 4 Argon's features, pricing and API.
FAQ
Is Kimi K3.1 Pro really open source?
Moonshot calls it open-weight and says the full weights will be released by November 4, 2026. It hasn't announced the license, and K3's license wasn't an OSI-approved open-source license.
How much does Kimi K3.1 Pro cost?
On the Kimi API, $2.00 per million input tokens, $10.00 per million output tokens and $0.20 per million cached input tokens. Highspeed costs $4.00 and $20.00.
Can it actually beat Claude Opus 5.5?
Not across the board. On Moonshot's numbers it's 0.7 points behind on Terminal-Bench 4.0 and ahead on DeepSWE and SWE-Marathon, but those results haven't been independently replicated yet.
Sources
Introducing Kimi K3.1 Pro, Moonshot AI (Kimi blog)
MoonshotAI: Kimi K3.1 Pro, OpenRouter
Kimi K3.1 Pro: specs, price, open weights date and benchmarks, Yotta Labs
Follow TheTechLane
Keep learning AI, development, and modern developer technologies with TheTechLane. Learn • Build • Share Follow @thetechlanehq for more developer guides, tutorials, and practical technology content.